Research library

Strategy families, and the honest limits of each

There is no universal best strategy, and anyone who names one is selling certainty rather than evidence. What the literature does support is that certain families of market behaviour have been documented repeatedly, across assets and decades — and that each of them fails in identifiable conditions. This library is organised around those families, with the failure modes written down beside the evidence.

Written by the Capstone Metals research desk. Reviewed by Daniel Kenney, President. Last updated February 2026.

Why families rather than a single system

A strategy that has worked in one regime is not evidence about the next one. Grouping ideas into families lets us ask a more useful question than "does this work?" — namely, "in which conditions has this behaved, in which has it broken, and what would combining it with something unlike it do?" Portfolio combinations across families are part of the research, not an afterthought.

The families, in the order we are building them

1. Time-series momentum / trend

Documented evidence: Persistence of an instrument's own past return over 1–12 month horizons has been documented across liquid futures markets (Moskowitz, Ooi & Pedersen, 2012).

What it does not prove: Documented in a sample period, before costs at the size we would trade, and with long flat or negative stretches. It is not a prediction.

Status: implemented —

2. Cross-sectional momentum

Documented evidence: Relative-strength ranking effects appear across and within asset classes (Asness, Moskowitz & Pedersen, 2013).

What it does not prove: Crowding, crash risk in reversals, and sensitivity to the ranking window and rebalance frequency.

Status: implemented —

3. Volatility-aware risk management

Documented evidence: Scaling exposure by recent realised volatility has been studied as a portfolio management overlay (Moreira & Muir, 2017).

What it does not prove: Reduces volatility, not loss. Adds turnover and therefore cost.

Status: implemented —

4. Value / fundamental factors

Documented evidence: Valuation spreads have explanatory power in cross-asset studies.

What it does not prove: Needs point-in-time fundamentals with reporting lags we do not yet ingest.

Status: research module —

5. Mean reversion / statistical arbitrage

Documented evidence: Short-horizon reversal and cointegration effects are widely studied.

What it does not prove: Highly sensitive to costs and to regime breaks; needs careful capacity analysis.

Status: research module —

6. Event and macro

Documented evidence: Scheduled macro releases move metals and rate-sensitive assets.

What it does not prove: Needs vintage-accurate release data, which is a later ingest module.

Status: research module —

7. Options-implied signals

Documented evidence: Implied volatility surfaces and skew carry information.

What it does not prove: No options data source. Deliberately not built.

Status: deferred — Deliberately postponed until the data and simulation quality justify it.

8. Order flow / microstructure

Documented evidence: Order-flow imbalance explains short-horizon price moves (Cont, Kukanov & Stoikov, 2014).

What it does not prove: Requires book/trade data and realistic execution modelling. Deliberately not built — a fabricated feature set would be worse than none.

Status: deferred — Deliberately postponed until the data and simulation quality justify it.

9. Reinforcement learning

Documented evidence: Promising in simulation.

What it does not prove: Only meaningful once the simulator and data discipline are mature. Deferred.

Status: deferred — Deliberately postponed until the data and simulation quality justify it.

How machine learning is allowed in

Machine learning enters as a layer on top of the research harness, not as a replacement for it: classifying market regimes, ranking across a cross-section, calibrating probabilities, selecting features, modelling residuals, spotting anomalies and weighting an ensemble. An unconstrained model that predicts price directly is not permitted to bypass the same walk-forward, holdout and cost tests as everything else. Reinforcement learning and larger neural models stay switched off until the data and the simulation are good enough to make them meaningful rather than impressive.

What a specification must contain

A version, a universe rule and a feature-set version.
An execution model, so the delay between signal and trade is explicit.
A cost model: commission, spread and the stress multiple used to re-test it.
A rebalance rule, position caps and risk limits.
The regimes it is claimed to apply to — and those it is not.
The evidence it came from, by source.

Evidence and sources

1. Moskowitz, Ooi & Pedersen, Time Series Momentum (2012) — The original cross-asset evidence for trend persistence at 1–12 month horizons.
2. Jegadeesh & Titman, Returns to Buying Winners and Selling Losers (1993) — The founding cross-sectional momentum result, and the reason the effect is studied at all.
3. Moreira & Muir, Volatility-Managed Portfolios (2017) — Scaling exposure inversely to recent volatility as a portfolio overlay rather than a signal.
4. Fama & French, A Five-Factor Asset Pricing Model (2015) — Value, profitability and investment as documented cross-sectional characteristics.
5. Novy-Marx & Velikov, A Taxonomy of Anomalies and Their Trading Costs (2016) — How trading costs erode each family, which is why cost assumptions are part of our specification.

Where this fits in the wider picture

Market Intelligence is the evidence layer. The wealth-protection side of Capstone is where those findings meet an actual plan — metals, retirement accounts and stewardship of what you already hold.

The education library — money, debt and purchasing power, from first principles.
Gold's place in the world economy — what the historical record does and does not show.
Wealth protection in one place — how metals, advice and entities fit together.

Research and education only. Nothing on this page is investment advice, a recommendation to buy or sell any security or metal, or a forecast. No outcome is promised or implied. Simulated and historical results do not indicate future results, and any strategy discussed here may lose money. Speak with us about your own circumstances before acting: (800) 200-9553.